Intel Skylake / Kaby Lake

Page 402 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Sweepr

Diamond Member
May 12, 2006
5,148
1,142
131
And also there are two entries which seems to be Coffee Lake based:

Genuine Intel(R) CPU 0000 @ 3.50GHz (6C 3.5GHz, 6x 256kB L2, 9MB L3)
http://ranker.sisoftware.net/show_r...d5e3dbecdfe6d4f280bd8dabceab96a680f3cef6&l=en

Genuine Intel(R) CPU 0000 @ 3.70GHz (4C 8T 3.7GHz, 4x 256kB L2, 6MB L3)
http://ranker.sisoftware.net/show_r...d5e3d0e2dbe2d7f183be8ea8cda895a583f0cdf5&l=en

Sisoft says it's a Kabylake Client platform but given that Kabylake doesn't support 6C this is more likely a misread. One interesting aspect, Sisoft reports 9MB L3 for the 6C version. It's a 6/6 version which might be Core i5 labeled and Core i7 might get 12 MB then. The 4C version however reports 6MB, so it also could mean Coffee Lake gets 1.5MB L3 per core. Or it's a misread, not sure about that. With only 1.5 MB per core the published Die size from benchlife seems high.

Intel(R) CofeeLake Mobile Graphics Controller
http://ranker.sisoftware.net/show_r...e7d6e3d0e6d4e7c1b38ebe98fd98a595b3c0fdc5&l=en

Some juicy info from this SiSoftware entry:

- 6C/6T Coffee Lake-S model(s) in the works, something I (personally) didn't expect - Intel seems to be taking advantage of the new hexa-core die
- Could end up as the new unlocked Core i5 - either way significant step up from their current designs at $200-300
- If they have a HT disabled version up and running at 3.5 GHz, full-fledged retail 6C/12T Core i7 should be clocked higher than that
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
The descriptions are interesting. 5x8 means 5 subslices of 8 EUs, which is same configuration as Gen 8 and 9. So that does not change.

5x8 = 40EUs, 8 ROPs, 20 TMUs


The page also says Gen10 GT2 features two slices.


#define GEN10_FEATURES(_gt, _slices, _l3)
+static const struct gen_device_info gen_device_info_cnl_gt2 = {
+ GEN10_FEATURES(2, 2, 6)
 
  • Like
Reactions: Dresdenboy

cbn

Lifer
Mar 27, 2009
12,968
221
106
- 6C/6T Coffee Lake-S model(s) in the works, something I (personally) didn't expect - Intel seems to be taking advantage of the new hexa-core die

Glad to hear that.

I also hope for a Coffee Lake X using that die (and thread count) at higher TDP.
 

SAAA

Senior member
May 14, 2014
541
126
116
And also there are two entries which seems to be Coffee Lake based:

Genuine Intel(R) CPU 0000 @ 3.50GHz (6C 3.5GHz, 6x 256kB L2, 9MB L3)
http://ranker.sisoftware.net/show_r...d5e3dbecdfe6d4f280bd8dabceab96a680f3cef6&l=en

Genuine Intel(R) CPU 0000 @ 3.70GHz (4C 8T 3.7GHz, 4x 256kB L2, 6MB L3)
http://ranker.sisoftware.net/show_r...d5e3d0e2dbe2d7f183be8ea8cda895a583f0cdf5&l=en

Sisoft says it's a Kabylake Client platform but given that Kabylake doesn't support 6C this is more likely a misread. One interesting aspect, Sisoft reports 9MB L3 for the 6C version. It's a 6/6 version which might be Core i5 labeled and Core i7 might get 12 MB then. The 4C version however reports 6MB, so it also could mean Coffee Lake gets 1.5MB L3 per core. Or it's a misread, not sure about that. With only 1.5 MB per core the published Die size from benchlife seems high.

Intel(R) CofeeLake Mobile Graphics Controller
http://ranker.sisoftware.net/show_r...e7d6e3d0e6d4e7c1b38ebe98fd98a595b3c0fdc5&l=en

The 4C-8T score is too old to be Coffe Lake based, it says 13 May 2016... but the first one looks to be the real deal.
Actually it's quite refreshing to see Coffe only 200 MHz slower than early Kaby Lake samples while sporting 2 more cores, I can't but expect the 7700K successor to stay at 4GHz base at least and possibly reach higher turbos thanks to 14nm++.
Overclocks will be entirely limited from temperature, delidding might be the only option to remove the excess heat, then six cores at 5GHz should be as easy as with Kaby Lake, if not guaranteed for this processor.

L3 Cache isn't really big: https://www.techpowerup.com/img/15-08-18/77a.jpg

How big is one SKL/KBL core? 10 mm²? 8 MB L3 shouldn't bigger than 5-6 mm², adding 4 MB more L3 would be tiny overall. CFL 6+2 is 149 mm² big according to Benchlife.

AMD slide at ISSCC 2017 stated that the 8 MB L3 cache is 19.1 mm2 for Skylake

The die shot points at some really small size for L3 indeed (pixel counting the green area gives me ~5mm2), but I'm left wondering WTH that yellow space around the blu cores and green LLC is then. Unused, dark silicon to cool off the die? But it has patterns and looks like pools of SRAM...

If it is part of the L3 it's easy to understand AMD claim of 19mm2 (I still get a bit less, ~16mm2 but whatever).
 
  • Like
Reactions: w3rd

itsmydamnation

Platinum Member
Feb 6, 2011
2,743
3,075
136
If this is true the slide I posted is way off. Not sure if I trust AMD numbers though.

We dont know if the intel number is skylake or broadwell. Also they likely include interfaces and L3 control logic that you dont see just by measuring the arrays.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
The page also says Gen10 GT2 features two slices.


#define GEN10_FEATURES(_gt, _slices, _l3)
+static const struct gen_device_info gen_device_info_cnl_gt2 = {
+ GEN10_FEATURES(2, 2, 6)

If the 5x8 description is true though, it excludes the possibility of having it a 2 slice device. Because Gen has subslices 5 doesn't divide into a whole number. Unless it has 5 slices of course.
 

Sweepr

Diamond Member
May 12, 2006
5,148
1,142
131
The 4C-8T score is too old to be Coffe Lake based, it says 13 May 2016... but the first one looks to be the real deal.

Nice catch. So what we have is a 6C/6T 3.5 GHz 9MB L3 entry, quite possibly from a new Core i5, hinting at 12MB L3 for the fully enabled parts.
 
Last edited:
  • Like
Reactions: Drazick

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
If the 5x8 description is true though, it excludes the possibility of having it a 2 slice device. Because Gen has subslices 5 doesn't divide into a whole number. Unless it has 5 slices of course.


The first digit doesn't reflect the numbers of subslices in one slice by the looks of it because even for the GT3 and GT4 it wasn't for Cannonlake: https://forums.anandtech.com/thread...-update-vr-zone.2418879/page-10#post-37185583

GT3 is 9x8 and GT4 13x8, they don't have just one slice for sure. The first digit must refer to the total count of subslices.

It still could have 6x8 physically on Die with one subslice being disabled on GT2, this would explain the 48 EUs on Sisoft. If some EUs or one subslice is faulty Intel could still use such a Die.
 

tamz_msc

Diamond Member
Jan 5, 2017
3,710
3,554
136
It's more likely that 6C+9MB L3 is what they can fit in 149 mm^2, rather than 6C+12MB L3.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
1.5MB per Core is expected for SKUs without HT, 9 MB is exactly what was expected for 6/6 models. It doesn't mean there are 12 MB available for 6/12 though, it's speculation at this point.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
Skylake is 12.25mm^2/core. The difference between 4+2 and 6+2 as posted here is 23 mm^2. That's why.


This is speculation at this point. Intel could reduce the dead space, it would easily fit into 149 mm². The difference between 9 MB and 12 MB is so tiny that we only could speculate for now.
 

tamz_msc

Diamond Member
Jan 5, 2017
3,710
3,554
136
77a.jpg


Not much dead space in Skylake 4+2. With a few tweaks Coffee Lake 6C may fit in 149 mm^2 with 9MB L3, 12MB is going to be much more difficult to fit on the same die size.
 
  • Like
Reactions: Drazick

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
The first digit doesn't reflect the numbers of subslices in one slice by the looks of it because even for the GT3 and GT4 it wasn't for Cannonlake: https://forums.anandtech.com/thread...-update-vr-zone.2418879/page-10#post-37185583

GT3 is 9x8 and GT4 13x8, they don't have just one slice for sure. The first digit must refer to the total count of subslices.

It still could have 6x8 physically on Die with one subslice being disabled on GT2, this would explain the 48 EUs on Sisoft. If some EUs or one subslice is faulty Intel could still use such a Die.

Interesting.

They've never done that though. If it's true they would be doing it for all configurations including GT4(which is likely nonexistent at this point but just for the sake of discussion). In fact such die disabled configurations only existed for products under GT1, before they replaced that with Atom based ones.

They used to, but that's when there was literally one die available for the entire segment. Now they are capable enough to create multiple different dies likely to optimize cost.

I do like the explanation that extra slices add different amount than the base one.
 
Last edited:

Sweepr

Diamond Member
May 12, 2006
5,148
1,142
131
DigiTimes reaffirms Skylake-X / Kaby Lake-X in June, Coffee Lake-S in August:

DigiTimes said:
As Intel's new Kaby Lake processors have not stimulated as much demand as expected since their launch in January, the sources are concerned that Intel's new top-end platform for June and new Coffee Lake processors for August may also face weak demand.

Another bit:

DigiTimes said:
Sources from motherboard players pointed out that notebooks have been gradually taking demand away from traditional PCs as a result of their better specifications, smaller form factors and cheaper prices.

www.digitimes.com/news/a20170424PD201.html
 

PaulIntellini

Member
Jun 2, 2015
58
4
71
The die shot points at some really small size for L3 indeed (pixel counting the green area gives me ~5mm2), but I'm left wondering WTH that yellow space around the blu cores and green LLC is then. Unused, dark silicon to cool off the die? But it has patterns and looks like pools of SRAM...

If it is part of the L3 it's easy to understand AMD claim of 19mm2 (I still get a bit less, ~16mm2 but whatever).

Yes, the yellow stuff is the LLC cache arrays. the "green" stuff is just the LLC/ring logic I think.
 
  • Like
Reactions: Ajay

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136

If they mean traditional PCs by Desktops as a proportion of sales, then no. The ratio is same as it was in the early 2000's. Nowadays, increasingly notebooks are replace-every-year $1000+ devices. Yes, there are a category of people that prefer Laptops. I think the proportion would have been greater, but Tablets took most of that away.
 

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
If they mean traditional PCs by Desktops as a proportion of sales, then no. The ratio is same as it was in the early 2000's. Nowadays, increasingly notebooks are replace-every-year $1000+ devices. Yes, there are a category of people that prefer Laptops. I think the proportion would have been greater, but Tablets took most of that away.
DIY only:
Despite some research firms believing that worldwide PC shipments' decline has already started slowing down, many motherboard players still see weak demand from the PC DIY channel, especially in China. As a result, the players believe their motherboard and graphics card shipments in 2017 are unlikely to recover and the overall volume may drop over 10% on year.